How fitness functions can help us govern and measure AI
Mar 6, 2025
auto_awesome
Rebecca Parsons, former CTO emerita of ThoughtWorks and co-author of 'Building Evolutionary Architectures', joins Neal Ford, a regular host and also co-author, to dive into the dynamic world of AI governance. They explore how fitness functions can optimize AI performance, ensuring systems meet their intended goals. The duo discusses identifying biases within AI, the importance of operationalizing large language models, and the need for objective metrics in rapidly changing tech landscapes. Their insights reveal how adaptability can shape the future of AI.
Fitness functions provide objective metrics that help organizations measure AI system goals, ensuring they meet defined performance criteria.
Collaborative ownership of fitness functions between architects and developers fosters shared responsibility, enhancing the effectiveness and relevance of AI systems.
Deep dives
Understanding Fitness Functions
Fitness functions are key metrics derived from evolutionary computation principles used to optimize complex systems. They provide a clear definition of desirable outcomes, allowing for the evaluation of various solutions to specific problems, such as minimizing distances in the traveling salesman problem. An important characteristic of these functions is their objectivity, which ensures that everyone agrees on whether a particular outcome meets the defined criteria. The ability to quantify vague concepts like maintainability into measurable fitness functions enables teams to automate checks, ultimately freeing them to focus on more nuanced architectural discussions.
AI and Fitness Functions
The discussion around AI emphasizes the need for fitness functions to help manage expectations on behavioral characteristics during rapid technological advancements. For instance, monitoring costs and latency can help organizations maintain control over project expenses and user experience, especially when utilizing generative AI models. A notable example discussed is the importance of latency in user interactions, where delays could lead to user abandonment of applications. Therefore, incorporating fitness functions to establish acceptable thresholds for these measures empowers teams to adapt architecture without compromising essential qualities.
Collaborative Ownership of Fitness Functions
The ownership of fitness functions is a collaborative effort between architects and developers, promoting shared responsibility for system performance. Architects typically design the fitness functions, while developers interact with them during the testing phase, contributing to their effectiveness. It is vital for developers to understand these functions as they serve as safeguards against significant architectural failures, making it important for development teams to share insights and raise issues. This cooperative approach ensures that fitness functions are relevant and tailored to the specific context of the applications being built.
Addressing Bias and Hallucinations with Fitness Functions
Fitness functions can play a crucial role in mitigating issues of bias and hallucinations in AI systems. For example, they can be designed to evaluate credit scoring algorithms to ensure that outputs remain unbiased across different demographic profiles. Regarding hallucinations, fitness functions can verify the credibility of information by checking against reliable sources, thereby enhancing the integrity of model outputs. This structured approach to defining potential biases prompts teams to think critically about objectives in AI deployment, ultimately fostering a more responsible use of technology.
AI is inherently dynamic: that's true in terms of the field itself, and at a much lower level too — models are trained on new data and algorithms adapt and change to new circumstances and information. That's part of its power and what makes it so exciting, but from a business and organizational perspective, that can make governance and measurement exceptionally difficult. How can we know that our AI is optimized for the right thing? How can we be sure it's oriented towards what we want it to be?
This is where the concept of fitness functions can help. Broadly speaking, fitness functions are ways of measuring the extent to which a given solution is fulfilling its goals — so, in the context of AI, they can help teams ensure that AI systems are serving their intended purpose.
In this episode of the Technology Podcast, Rebecca Parsons and Neal Ford — authors (alongside Pat Kua and Pramod Sadalage) of Building Evolutionary Architectures, the book which brought fitness functions into the software architecture space — join host Ken Mugrage to explore how the fitness function concept can help us better manage the dynamism of AI and, in doing so, overcome the challenge of bringing such systems into production.